DenseChip

GPU Design · All levels

Compute Fabric & Memory System: Tricky Q&A

Senior interview and review questions for Compute Fabric & Memory System.

Section Q&A bank

Use these drills after completing all topics in Compute Fabric & Memory System. Answer with workload context, mechanism proof, artifact, owner, and release decision.

How do you distinguish NoC congestion from memory-controller saturation?

diagram
[INT][GPU][COMPUTE-INTERCONNECT]

Q: How do you distinguish NoC congestion from memory-controller saturation?

A:
NoC issues show hop-level hotspot imbalance, while controller saturation shows queue depth and bank-policy stalls near memory interfaces.

FOLLOW-UP TRAP: Using only end-to-end latency to classify bottlenecks.

Why can peak-link bandwidth still under-deliver workload throughput?

diagram
[INT][GPU][COMPUTE-INTERCONNECT]

Q: Why can peak-link bandwidth still under-deliver workload throughput?

A:
Traffic patterns, arbitration fairness, and burst alignment can waste available link capacity despite high theoretical limits.

FOLLOW-UP TRAP: Equating link spec sheets with sustained application behavior.

What is the first audit for host-device transfer regressions?

diagram
[INT][GPU][COMPUTE-INTERCONNECT]

Q: What is the first audit for host-device transfer regressions?

A:
Check queue depth, transfer granularity, and overlap with kernel execution before changing PCIe or firmware policy.

FOLLOW-UP TRAP: Jumping to hardware blame without DMA timeline evidence.

When should fabric topology be reconsidered in roadmap planning?

diagram
[INT][GPU][COMPUTE-INTERCONNECT]

Q: When should fabric topology be reconsidered in roadmap planning?

A:
When repeated workloads hit the same congestion classes despite scheduling and QoS policy tuning.

FOLLOW-UP TRAP: Treating chronic topology limits as temporary software issues.

Q&A drill guide

diagram
WORKLOAD -> BOTTLENECK -> METRIC -> CHANGE -> REGRESSION

Sketch while answering

diagram
COMPUTE FABRIC

SM clusters <-> NoC <-> L2 <-> memory controllers <-> HBM

Key takeaways

  • Always pair micro-kernel metrics with end-to-end workload impact.

  • Lock toolchain, driver, and launch metadata before comparing performance results.

Common pitfalls

  • Optimizing occupancy without checking memory-system saturation.

  • Comparing profiler captures from different driver or compiler builds.

  • Declaring wins without reproducible accuracy and performance gates.

Practice this lesson